[nvq++] Remove kernel bodies in fixup-linkage

[lmondada]

Currently, in MLIR mode kernel bodies are compiled by clang but never called: the MLIR-compiled kernels always shadow the LLVM-compiled ones. This PR removes these kernel bodies in the LLVM IR, when in MLIR mode. This will speed-up compilation and (in MLIR mode) will allow us to remove definitions of C++ types and functions that are only used within kernels (e.g. cudaq::qvector).

This PR is made of 2 commits, which can be reviewed independently:

• the first commit rewrites the fixup-linkage script using LLVM instead of ad-hoc string/regexp parsing of .ll files. The rewritten script is equivalent to the original
• the second commit uses the parsed LLVM module to replace the body of each kernel function with a body made of a single unreachable op. If this op is ever reached, the program would crash, but because the program control flow is always redirected to the MLIR kernels, this can never happen.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[1tnguyen]

👍